Liquid AI published LFM2.5-24B-A2B on Hugging Face on October 13, 2026, describing it as a 24-billion-parameter reasoning model with 2 billion active parameters per token, built for what it calls fast, private agents that run on local hardware. According to the post, the model builds on LFM2-24B-A2B, keeps a mixture-of-experts architecture of 40 layers with 64 experts and top-4 routing, and expands the context window from 32K to 131K tokens. Pre-training was scaled from roughly 17 trillion to about 33 trillion tokens. The company says it is a reasoning-only model that produces an explicit chain of thought before answering, and that it calls tools and handles multi-step tasks inside agent harnesses. Liquid AI reports it evaluated the model against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning across 16 benchmarks, stating it leads or matches on five of eight tool-use benchmarks and all three instruction-following benchmarks, while larger models retain an edge on knowledge-heavy and coding tasks. The post cites edge decoding speeds of 184 tokens/s on an M5 Max and 114 tokens/s on a Ryzen AI Max+ 395. These figures come from the vendor's own announcement and have not been independently verified.
- 24B total parameters with 2B active per token; context window expanded to 131K tokens
- Vendor reports leading or matching on five of eight tool-use benchmarks against larger models
- Claimed edge speeds of 184 tokens/s on M5 Max and 114 tokens/s on Ryzen AI Max+ 395
- Day-one support cited for llama.cpp, MLX, vLLM, SGLang and ONNX; published October 13, 2026
What it means for you
This is a model designed to run agents on your own computers rather than through a cloud API, which matters if you handle data you'd rather not send off-site. The headline claim is that it stays fast and cheap while running locally. The benchmark numbers come from the company itself, so treat them as claims to test, not proven facts.
Try this
If you already have someone technical running local models, have them download it and run it against one real task you care about — ideally something involving your own sensitive data — and compare the result to whatever tool you use now.
Who should care
Technical teams already experimenting with on-device or self-hosted AI, and organizations with data-privacy constraints that rule out cloud APIs for certain workloads.
Skip this if
You have no one running models locally, no privacy reason to avoid cloud tools, or you're not already building agent workflows — this changes nothing for you today.
Sources: Hugging Face — read the original